[SYCL] Add support for __arithmetic_fence builtin for SYCL targets.

[zahiraam]

__arithmetic_fence enforces ordering on expression evaluation when fast-math is enabled.
In fast math mode some floating-point optimizations are performed such as reassociation and distribution.

For example, the compiler may transform (a+b)+c into a+(b+c). Although these two expressions are equivalent in integer arithmetic, they may not be in floating-point arithmetic. The builtin tells the compiler that the expression in parenthesis can’t be re-associated or distributed.
__arithmetic_fence(a+b)+c is not equivalent to a+(b+c).

[Fznamznon]

One NIT, otherwise lgtm

[zahiraam]

@intel/llvm-gatekeepers I don't think the failure is related to this PR. In which case the PR is ready to merge. Thanks.

[elizabethandrews]

Is there a reason we are doing this in intel/llvm and not community?

[zahiraam]

Is there a reason we are doing this in intel/llvm and not community?

I can't think of any reason.

[elizabethandrews]

Is there a reason we are doing this in intel/llvm and not community?

I can't think of any reason.

I would suggest attempting to put this in community instead then. We always try to do that if possible.

[zahiraam]

Is there a reason we are doing this in intel/llvm and not community?

I can't think of any reason.

I would suggest attempting to put this in community instead then. We always try to do that if possible.

I just verified and sycl_device attribute is not implemented.

[bader]

Is there a reason we are doing this in intel/llvm and not community?

I can't think of any reason.

I would suggest attempting to put this in community instead then. We always try to do that if possible.

I just verified and sycl_device attribute is not implemented.

sycl_kernel is supported.

[bader]

@zahiraam, could you add tag(s) for the PR title, please? https://github.com/intel/llvm/blob/sycl/CONTRIBUTING.md#pull-request

Add support for __arithmetic_fence builtin for SYCL targets.

SYCL or SPIR?

[bader]

I suggest following @elizabethandrews suggestion and commit this change to llvm.org.

A couple of notes:

• I don't think this change is SYCL specific and we need to replace SYCL with SPIR in the log message.
• Test doesn't need to use SYCL mode as well. You can keep using C++, just add compilation for the SPIR target.
  • If you'd like to check this feature in SYCL mode, please, add new SYCL specific tests in CodeGenSYCL and SemaSYCL directories. You can use existing SYCL tests as a reference.